We Built the Same App for $1.16. Claude Code Spent $10.89 — and Still Needed a Fix.
A side-by-side cost benchmark: Empromptu's Builder vs. Claude Code building an identical app from the same spec. One got it right the first time for just over a dollar. The other needed a second pass and cost 9.4x more.
The experiment
We gave two systems the same brief: build "Haiku Twitter." A user types a topic, gets an AI-generated haiku, shares it to a public feed, likes other posts, and can remix any poem into a new one. Accounts required. No moderation. No profile pages. A clean, zen visual theme. Nothing exotic — the kind of scoped, real-world app a customer actually asks for.
One build ran on Empromptu's Builder, using Sonnet 5. The other ran on Claude Code, using Opus 4.8 — a larger, more expensive model.
The numbers
Empromptu's Builder used 442,880 input tokens, 129,624 cache-read tokens, 11,152 cache-write tokens, and 22,250 output tokens. At Sonnet 5 rates, that's $1.16 total. The app worked correctly on the first try.
Claude Code used 1,504 input tokens, 10,634,356 cache-read tokens, 336,760 cache-write tokens, and 26,694 output tokens on its first pass — $8.10 at Opus 4.8 rates. That first build had a bug: environment variables weren't configured correctly. Fixing it took another 364 input tokens, 1,086,265 cache-read tokens, 341,867 cache-write tokens, and 4,352 output tokens — $2.79 more.
Total cost for Claude Code: $10.89. Same app. Nearly 10x the price — and it still took two tries to get there.
Why the gap is this big
The difference isn't model intelligence. It's what each system carries into every step.
Empromptu's Builder continuously summarizes as it works, and doesn't cache ephemeral context. That means a few more raw input tokens go in, but far fewer cache writes accumulate. The system never has to drag a growing pile of history forward through the whole build — which is exactly what let it use a smaller, faster model and still get the app right the first time.
Claude Code takes the opposite approach: it caches everything, and re-introduces all of it at every step. Look at the cache-read numbers — over 10.6 million tokens on the first build alone. That's not "more context," that's context rot: the model's reliability gets shakier as irrelevant history piles up around the part of the task that actually matters. Even running a larger, more capable model, Claude Code still needed a second pass to fix something Empromptu's Builder didn't get wrong at all.
The takeaway
This is one build, not a claim that holds for every possible app — but it's a clean, apples-to-apples look at what architecture costs in practice. Avoiding context rot isn't only a stability win. Here, it was a 9.4x cost difference, because a cleaner context is what let us stop paying a premium for a bigger model to compensate for a messier one.